Load Balancing

1. A method for load balancing Ethernet traffic within a fat tree network comprising: randomly assigning incoming messages into hash classes using a hash function; programming the assignment of said hash classes among uplinks based on data center wide load balancing measurements; and transmitting said incoming messages on said uplinks according to said hash class.

2. The method of claim 1 , further comprising dynamically allocating said hash classes among uplinks in response to a measurement of network traffic.

3. The method of claim 1 , further comprising: initializing said fat tree network using a spanning tree algorithm to generate a spanning tree network comprising spanning tree links; and then activating previously disconnected core switches which create redundant links.

4. The method of claim 3 , further comprising flooding information through said fat tree topology using said spanning tree links.

5. The method of claim 4 , further comprising flooding information through only the original spanning tree structure.

6. The method of claim 3 , in which the core switches are turned off while generating a spanning tree network.

7. The method of claim 3 , in which the initialized fat tree network has a single connection to each of the edge switches.

8. The method of claim 3 , in which activating core switches which create redundant links comprises turning on remaining core switches.

9. The method of claim 1 , further comprising: gathering hash statistics across said fat tree network; and reporting said hash statistics to a central controller.

10. The method of claim 9 , further comprising: evaluating said hash statistics using said central controller; and dynamically setting an allocation of said hash classes among said uplinks to balance said Ethernet traffic within said fat tree network.

11. The method of claim 1 , further comprising setting said allocation of said hash classes among said uplinks such that individual flows follow the same outbound and inbound paths.

12. The method of claim 1 , further comprising allocating said hash classes among said uplinks is performed according to a greedy heuristic algorithm.

13. The method of claim 1 , further comprising allocating said hash classes among said uplinks is performed according to a random algorithm.

14. The method of claim 1 , further comprising down linking messages from a root switch to an edge switch is performed using an adaptive forwarding technique.

15. The method of claim 1 , further comprising: statically balancing loads within said fat tree network; and dynamically balancing loads within said fat tree network.

16. The method of claim 1 , in which the hash class is assigned based on a destination address of an incoming message.

17. A fat tree network having load balancing of communication flows comprises: end stations; enhanced edge switches; said enhanced edge switches having downlinks to said end stations and using a hash module to categorize incoming messages into hash classes; and commodity core switches; said enhanced edge switches having uplinks to said commodity core switches, said hash classes being dynamically allocated to said uplinks to load balance said communication flows within said fat tree network based on network traffic information gathered across said fat tree network.

18. An enhanced network switch for load balancing communication flows in a fat tree network comprising: downlinks; uplinks, said enhanced network switch being configured to: route communication flows among said downlinks and uplinks; receive a MAC address from a message; and output a hash address; a TCAM lookup module; said TCAM lookup module allocating said hash address into a hash class and allocating said hash class to one of said uplinks; a forwarding cache to receive said hash address and look up a corresponding destination port; wherein one of said uplinks is a designated link, said designated link being a link with an underlying spanning tree network; if said hash address does not correspond to said destination port, then said message is flooded preferentially on said designated link.

19. The network switch of claim 18 , further comprising: a central controller; and wherein said enhanced network switch gathers hash statistics which measure the load balance among said uplinks and reporting said hash statistics to said central controller; wherein said central controller dynamically sets an allocation of said hash classes among said uplinks.

20. The network switch of claim 18 , in which the hash address is output from a symmetric hash function.